Research map / Vision-language and multimodal / CLIP and zero-shot
DeCap: Decoding CLIP Latents for Zero-Shot Captioning via Text-Only Training
Wei Li, Linchao Zhu, Longyin Wen, Yi Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Project the image embedding into the text embedding space instead of using the visual embedding directly (ablation) | DeCap · MSCOCO, in-domain unpaired | CIDEr: 25.7 -> 91.2 | Improved |
| Decode the nearest text embedding rather than retrieve its original caption (ablation) | CLIPRe / DeCap-NND · MSCOCO, in-domain unpaired | CIDEr: 53.4 -> 62.9 | Improved |
| Train the decoder on SS1M rather than CC3M-text | DeCap · MSCOCO Karpathy-test, zero-shot | CIDEr: 42.1 -> 50.6 | Improved |
| Add target-domain text to construct the support memory (ablation) | DeCap · Flickr30K to MSCOCO, cross-domain | CIDEr: 44.4 -> 63.1 | Improved |
| Train the decoder on the full MSCOCO text set instead of 1% of it (ablation) | DeCap · MSCOCO Karpathy-test | CIDEr: 81.5 -> 91.2 | Improved |
| Use a similarity-filtered support memory instead of randomly sampling the same number of embeddings (ablation) | DeCap · MSCOCO Karpathy-test | CIDEr: 38.2 (-4.0) -> 42.3 (+0.1) | Improved |
| Use decoder-reconstructed captions as the support memory (ablation) | DeCap · MSCOCO Karpathy-test | CIDEr: 91.2 -> 95.1 (+3.9) | Improved |
| Add the prompt “Attention! There is/are” (ablation) | DeCap-BookCorpus · MSCOCO Karpathy split | CIDEr: 21.8 -> 31.9 (+10.1) | Improved |
| Add the prompt “Attention! There is/are” (ablation) | DeCap-CC3M · MSCOCO Karpathy split | CIDEr: 42.1 -> 42.2 (+0.1) | No clear effect |
| Fine-tune on MSCOCO text after pre-training on CC3M text (ablation) | DeCap · NoCaps val, overall | CIDEr: 38.3 -> 58.2 | Improved |
| Switch from CC3M pre-training to MSCOCO pre-training (ablation) | DeCap · NoCaps val, out-of-domain | CIDEr: 48.7 -> 25.8 | Hurt |
| Replace random support sampling with a 1% support memory (ablation) | DeCap · MSCOCO Karpathy-test | CIDEr: 3.8% performance drop in CIDEr | Hurt |
Later papers that compared with this method
6 later accepted papers report DeCap as a baseline in their results tables; 5 of them report a better number on most of the results they share with it.
- One Patch to Caption Them All: A Unified Zero-Shot Captioning Framework (CVPR 2026 · better on 2 of 2 shared results)
- Diffusion Bridge: Leveraging Diffusion Model to Reduce the Modality Gap Between Text and Vision for Zero-Shot Image Captioning (CVPR 2025 · better on 1 of 4 shared results)
- Image Captioning with Multi-Context Synthetic Data (AAAI 2024 · better on 7 of 8 shared results)
- Improving Cross-Modal Alignment with Synthetic Pairs for Text-Only Image Captioning (AAAI 2024 · better on 3 of 4 shared results)
- MeaCap: Memory-Augmented Zero-shot Image Captioning (CVPR 2024 · better on 3 of 3 shared results)
- Transferable Decoding with Visual Entities for Zero-Shot Image Captioning (ICCV 2023 · better on 1 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ZeroCap — ZeroCap: Zero-Shot Image-to-Text Generation for Visual-Semantic Arithmetic (CVPR 2022) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Mining Fine-Grained Image-Text Alignment for Zero-Shot Captioning via Text-Only Training (AAAI 2024 · similarity 0.92)
- Modeling Caption Diversity in Contrastive Vision-Language Pretraining (ICML 2024 · similarity 0.89)
- Zero-Shot Image Captioning with Multi-type Entity Representations (AAAI 2025 · similarity 0.89)
- SuperCLIP: CLIP with Simple Classification Supervision (NeurIPS 2025 · similarity 0.89)
- Interpreting and Analysing CLIP's Zero-Shot Image Classification via Mutual Knowledge (NeurIPS 2024 · similarity 0.88)
- TripletCLIP: Improving Compositional Reasoning of CLIP via Synthetic Vision-Language Negatives (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
